As an essential part of Internet of Things, monocular depth estimation predicts dense depth maps from a single RGB image captured by monocular cameras. Past monocular depth estimation methods almost focus on improving accuracy at the cost of increased latency, power consumption, and computational complexity, failing to balance accuracy and efficiency. Additionally, when speeding up depth estimation algorithms, researchers commonly ignore their adaptation to different hardware architectures on edge devices. This paper aims to solve these challenges. (1) We design an efficient monocular depth estimation model (MDE) for precise depth sensing on edge devices. (2) We employ a reinforcement learning algorithm and automatically prune redundant channels of MDE by finding a relatively optimal pruning policy. The pruning approach lowers model runtime and power consumption with little loss of accuracy through achieving a target pruning ratio. (3) We accelerate the pruned MDE, while adapting it to different hardware architectures with a compilation optimization method. The compilation optimization further reduces model runtime by an order of magnitude on hardware architectures. Extensive experiments confirm that our methods are effective for images of different sizes on two public datasets. The pruned and optimized MDE achieves promising depth sensing with a better trade-off among model runtime, accuracy, computational complexity, and power consumption than the state of the arts on different hardware architectures.

Given monocular images as inputs, monocular depth estimation (MDE) infers pixel-level depth. MDE is always a critical stage in scene sensing on edge devices. Existing MDE studies frequently employ deep neural networks (DNNs) for MDE, but they still face some problems, such as sacrificing computational complexity and efficiency in return for great precision, or losing more precision in exchange for increased efficiency. To alleviate these issues; 1) we propose an encoder–decoder network (EdgeNet) for precise and fast MDE on different edge devices. When recovering depth in the decoder, we design upsampling modules to aggregate global depth information with low computational complexity, improving the accuracy of the decoder by extracting its different ranges of depth information; 2) we develop a two-stage channel pruning method to, respectively, prune the encoder and decoder based on their characteristics. Our pruning method further reduces latency and model/computational complexity of EdgeNet, while losing little accuracy; and 3) we optimize the pruned EdgeNet to decrease graphics processing unit (GPU) scheduling overhead. The optimization accelerates MDE inference by an order of magnitude on the TX2 GPU device, when the input resolution is 224 $\times $ 224. Extensive experiments show that our strategies are effective on different edge GPU devices, when input resolutions differ in outdoor or indoor scenes. For example, compared with the state of the art, the optimized EdgeNet, respectively, reduces the GPU latency by 76.3% and 89.2% on Nano and TX2 GPU devices with 2.6% lower root mean square error when the input resolution is 128 $\times $ 416.

As overhead contact (OC) is an essential part of power supply systems in high-speed railways, it is necessary to regularly inspect and repair abnormal OC components. Relative to manual inspection, applying LiDAR (light detection and ranging) to OC inspection can improve efficiency, accuracy, and safety, but it faces challenges to efficiently and effectively segment LiDAR point cloud data and identify catenary components. Recent deep learning-based recognition methods are rarely employed to recognize OC components, because they have high computational complexity, while their accuracy needs to be improved. To track these problems, we first propose a lightweight model, RobotNet, with depthwise and pointwise convolutions and an attention module to recognize the point cloud. Second, we optimize RobotNet to accelerate its recognition speed on embedded devices using an existing compilation tool. Third, we design software to facilitate the visualization of point cloud data. Our software can not only display a large amount of point cloud data, but also visualize the details of OC components. Extensive experiments demonstrate that RobotNet recognizes OC components more accurately and efficiently than others. The inference speed of the optimized RobotNet increases by an order of magnitude. RobotNet has lower computational complexity than other studies. The visualization results also show that our recognition method is effective.

Depth estimation has received considerable attention and is often applied to visual simultaneous localization and mapping (SLAM) for scene reconstruction. At least to our knowledge, sufficiently reliable depth always fails to be provided for monocular depth estimation-based SLAM because new image features are rarely re-exploited effectively, local features are easily lost, and relative depth relationships among depth pixels are readily ignored in previous depth estimation methods. Based on inaccurate monocular depth estimation, SLAM still faces scale ambiguity problems. To accurately achieve scene reconstruction based on monocular depth estimation, this paper makes three contributions. (1) We design a depth estimation model (DEM), consisting of a precise encoder to re-exploit new features and a decoder to learn local features effectively. (2) We propose a loss function using the depth relationship of pixels to guide the training of DEM. (3) We design a modular SLAM system containing DEM, feature detection, descriptor computation, feature matching, pose prediction, keyframe extraction, loop closure detection, and pose-graph optimization for pixel-level scene reconstruction. Extensive experiments demonstrate that the DEM and DEM-based SLAM are effective. (1) Our DEM predicts more reliable depth than the state of the arts when inputs are RGB images, sparse depth, or the fusion of both on public datasets. (2) The DEM-based SLAM system achieves comparable accuracy as compared with well-known modular SLAM systems.

Detecting drones is a complex challenge, primarily due to their small feature size for extraction and variable lighting conditions. It is crucial to effectively capture and model features for drone detection. To accurately detect drones, we propose feature capture and modeling modules. The feature capture module has a minimal number of FLOPs and parameters. It consists of both local and global attention branches, which capture contextual information and global dependencies across the entire feature set. Complementing this, our feature modeling module innovatively calculates attention maps without any additional parameters. This module augments the capability of the feature capture mechanism to represent complex patterns more effectively. Finally, to ensure rapid deployment, we convert the proposed models to machine codes by introducing a compiler, accelerating drone detection. The compiler unifies inter- and intra-operator scheduling with task abstraction. It optimizes the codes for hardware. In compiling time, the effective schedule is performed. The compilation ensures that drone detection is real-time and accuracy remains unchanged. Through rigorous testing, our methods have demonstrated superiority over most current ones in several metrics, including accuracy, parameter quantities, FLOPs, average FPS, visual effects, and latency. Our method yields at least 23.5% and 12.47% higher AR_M than existing methods on DUT-Anti-UAV and Online Drone datasets. Our inference speed is at least 6.49% higher than other methods on NVIDIA RTX 2080 Ti GPU.